Why Do Training and Prediction Use Different Numbers of Data Points?

[gaosanyuan]

Maybe I didn't understand well.
In my opinion, the number of points used in training are determined by event_shape, which is default 256.
And in denosie eval, I found the number of points used in prediction is 1024.

                        if x_points.shape[0] > 1024:
                            indices = subsample_indices(n_points_orig, subsample_size=1024, seed=-1)
                            x_points = x_points[indices, :]
                            x_points_pp = x_points_pp[indices] if x_points_pp is not None else None
                            indices_not_selected = np.delete(np.arange(n_points_orig), indices)
                            x_points_mask_indices_not_selected = np.where(x_points_mask.detach().cpu().numpy())[0][
                                indices_not_selected]
                            x_points_mask[x_points_mask_indices_not_selected] = False

How to under the difference between them, Thanks